Local models

Run models on your own machine with Ollama, LM Studio, Atomic Chat or Anaconda Desktop.

Hull can use models that run on your own machine. Your prompts and code stay local, you can work offline, and there are no per-token charges.

The trade-offs:

Hardware. Good results need a capable GPU or a machine with plenty of unified memory.
Capability. Local models are smaller than frontier cloud models. They are more likely to loop, misuse tools or produce broken code.
Speed. Responses are often slower, especially with long contexts.
Features. Many local models don't support prompt caching, image input or reliable tool calling.

For agent work, choose a model that supports tool calling and set a context window of at least 32K tokens.

Supported local runtimes

Runtime	How Hull connects	Default address
Ollama	Custom OpenAI-compatible provider	`http://localhost:11434/v1`
LM Studio	Built-in `lmstudio` provider	`http://127.0.0.1:1234/v1`
Atomic Chat	Built-in `atomic-chat` provider with model discovery	`http://127.0.0.1:1337/v1`
Anaconda Desktop	Discovers and imports the running model server	Set by Anaconda Desktop

Any other server that exposes an OpenAI-compatible API (vLLM, llama.cpp server and others) works as an OpenAI-compatible provider.

Set limit.context and limit.output for local models so Hull can manage the context window. See Custom models.

Connection refused: the local server isn't running, or it's on a different host or port than Hull expects. Check the base URL against the table above.
Slow responses: try a smaller model or a shorter context.
Model not found: the model ID in Hull must match the ID the server reports. For Ollama, use the same name you pass to ollama run.
The model doesn't use tools: switch to a model that supports tool calling.